The Marcellus Shale formation covers an area of approximately 240 000 km2 across several states: New York, Pennsylvania, Ohio, West Virginia, Maryland, and Virginia (Kargbo et al., 2010) (Fig. S1). The Marcellus Shale Energy and Environment Laboratory (MSEEL) is an approximately 14 000 m2 study well pad in Morgantown, WV, USA (39.602∘ N, 79.976∘ W) (MSEEL, 2019). The MSEEL is a multi-institutional, long-term collaborative field site where integrated geoscience, engineering, and environmental research has been conducted to assess environmental impacts and develop new technology to improve recovery efficiency as well as reduce environmental footprint of shale gas operations (MSEEL, 2019). The MSEEL is the site of two horizontal production wells completed in 2011 (wells 4H and 6H, Fig. 1) and two horizontal production wells completed in 2015 (wells 3H and 5H, Fig. 1). Production from the newer horizontal wells began in December 2015. Figure 1 shows the location of the trailer with respect to the location of the wells and the boundaries of the well pad. The distance between the wells and the trailer is 90 m. Dates and duration for phases of operation are shown in Fig. S2, the total gas production for the four wells is shown in Fig. S3. The vertical drilling was conducted using three diesel Caterpillar 3512 engines with 1365 kW generators. Horizontal drilling made use of two dual fuel (40 % diesel and 60 % natural gas) engines. All activities at the well pad followed the industry's best management practices (MSEEL, 2019).

An ambient air monitoring laboratory (5.5 m trailer with ambient air sampled from inlets on the trailer roof) was situated at the northeastern corner of the MSEEL well pad (Fig. 1). With wind direction at this location most frequently from the southwest (Fig. 2), this position optimized the occurrences of the laboratory being downwind of the well pad. Instrumentation in the laboratory and measured constituents are listed in Table 1. All instruments were maintained and calibrated according to manufacturer's recommended protocols. Details of the laboratory assembly and operation have been previously described (Pekney et al., 2014).

Data collected at the air monitoring site are classified by activity at the well pad. Horizontal drilling occurred on 8 September–5 October 2015, first at well 5H then at well 3H. Hydraulic fracturing occurred on 10 October–16 November. Cleanout activities followed on 20–26 November, which involved using a diesel-powered coil tubing rig to drill out plugs and flush out residue left in the wells.

Flowback, the flowing of gas, formation fluid, and hydraulic fracturing fluid up the wells to the surface, took place over 10–14 December, after which both wells were in production. A reduced emission completion (REC) was performed; gas produced during this time was captured using portable equipment brought on site that separates the gas from the liquids so that the gas can be retained as a product.

Air monitoring began 18 September 2015 and ended 1 February 2016. No data were collected for the vertical drilling phase. Data collection was continuous except for calibration and instrument downtime. The laboratory's meteorological station measured relative humidity, temperature, rainfall, solar radiation, wind direction, wind speed, and barometric pressure at an elevation of 10 m.

Positive matrix factorization (PMF), a factor analysis method (Fig. S7), was applied to hourly averaged ambient concentrations of measured species to identify possible sources and patterns for the stages of development. PMF decomposes the sample data into two matrices: factor profiles (representative of sources) and factor contributions (Brown et al., 2015; Norris et al., 2014). The fundamental objective of PMF is to solve the chemical mass balance (Eq. 1) between measured species concentrations and source profiles while optimizing goodness of fit (Eq. 2).

Mass balance is as follows (Evans and Jeong, 2007): 1xi,j=∑k=1pgikfkj+eij, where xi,j is the data matrix with dimensions of i (observations) by j (chemical species), p is the optimal number of factors, gik is the factor contribution to the observation, fkj is the species profile of the factor, k is the factor, and ei,j is the residual concentration for each observation.

Goodness of fit is as follows: 2Q=∑i=1n∑j=1mxij-∑k=1pgikfijsij2, where Q is the goodness of fit, n is the total number of observations, m is the total number of chemical species, and sij is the uncertainty for each observation. Summary of methods for uncertainty calculations are provided in the Supplement. Missing values within the dataset are replaced with the median value of that species; also, uncertainty for missing values is set at 4 times the species-specific median by the program. Multiple runs with different numbers of factors are executed for each dataset. The output of the PMF analysis needs to be interpreted by the user to identify the number of factors that may be contributing to the samples and the possible sources they represent. One of the main strengths of PMF analysis is that each sample is weighted individually, which allows the user to adjust the influence of each sample based on the measurement confidence.

Signal-to-noise ratio (S/N), an indicator of the accuracy of the variability in the measurements, can be used to identify a species as “strong”, “weak”, or “bad”. Generally, if this ratio is greater than 0.5 but less than 1, that species has a “weak” signal. “Strong” is the default value for all species with an assumption of S/N greater than 1. The “bad” category excludes the species from the rest of the analysis. We considered the number of samples that are missing or below the detection limit when choosing the category for each species. (Norris et al., 2014). The expected goodness of fit (Qexpected) is calculated for each scenario (Norris et al., 2014).

Expected goodness of fit is as follows: 3Qexpected=i×j-p×i+p×j, where (i×j) is the number of non-weak data values in Xij, and (p×i) and (p×j) are the number of elements in G and F, respectively. Qrobust is the calculated goodness-of-fit parameter that excludes points that are not fit by the model. The lowest Qrobust/Qexpected is calculated to compare different factor scenarios; when changes in Q become small with increasing factors, it can indicate that there may be too many factors in the solution (Brown et al., 2015).

In addition to these calculated parameters, factor profiles and error estimation diagnostics are used to compare the output of different simulations. Marker species (chemical species that are unique to a particular source) and temporal or seasonal variations can be used to aid in identifying the possible emission sources (Fig. 3). Associations between factors can also provide useful information for profile characterization. Moreover, meteorological data can provide useful information about the geographic location of the sources.

In order to perform the PMF analysis, we utilized a user-friendly graphical user interface (GUI) developed by the U.S. Environmental Protection Agency (EPA), EPA PMF 5.0 (Norris et al., 2014). Hourly average data were used for each pollutant to unify the measurement intervals. All pollutants included in the matrix were identified as “strong” (signal-to-noise ratio: S/N >2). A total of 50 base runs were performed, and the run with the minimum Q value was selected as the base run solution. In each case, the model was run in the robust mode with a number of repeat runs to ensure the model least-squares solution represents a global rather than a local minimum. First, the rotational (linear transformation) Fpeak variable was held at the default value of 0.0. However, there can be almost infinite possibilities of F and G matrices that produce the same minimum Q value, but the goal is producing a unique solution. As a result, rotational freedom is one of the main sources of uncertainty in PMF solutions (Paatero et al., 2014). Therefore, Fpeak values were adjusted (-1.0, -0.5, 0.5, and 1.0) to explore how much rotational ambiguity exists in PMF solutions. In other words, the model adds and/or subtracts rows and columns of F and G matrices based on the Fpeak, which is typically between -5 and +5 (Norris et al., 2014). Positive Fpeak values cause a sharpened F matrix and smeared G matrix; negative Fpeak values result in subtractions in the G matrix. The factor contributions were analyzed to find the optimum Fpeak value.

The PMF analysis was completed with error estimation. We used three methods of error estimation: bootstrap (BS), displacement (DISP), and BS–DISP, which guide understanding the stability of the PMF solution (Norris et al., 2014). BS analysis is used to determine whether a set of observations affect the solution disproportionately. The main idea of BS analysis is to resample different versions of the original dataset and perform PMF analysis. Random errors and rotational ambiguity affect BS error intervals. The main reason for rotational ambiguity is the existence of infinite solutions similar to the solution generated by PMF solution. DISP analysis helps to analyze the PMF solution in detail. Only rotational ambiguity affects DISP error intervals.

BS–DISP is a hybrid method that gives more robust results than DISP results.

Figure 3 shows a box-and-whisker graph of the measured NOx, NO, NO2, ozone, and ethane during the whole monitoring campaign at the study site. Similarly, Fig. 5 shows a statistical summary of methane and carbon dioxide. The y axis represents concentrations and the x axis represents the phases of the well development. The black line on each of the boxes represents the median for that particular dataset. The small circles represent outliers. The blue circles represent the mean. Since most of the VOC concentrations measured were consistently below 10 ppb, only ethane is included. There was an increase in NOx (25th percentile (q1)=12.5 ppb) and NO (q1=2.7 ppb) during the fracturing phase compared to other phases. The whiskers show the high variability for this phase, which can be a result of small sample size for the fracturing phase. The NO/NO2 ratio for 25th and 75th percentiles was 1.2, indicating fresher, less oxidized emissions. The skewness of the data for this phase indicates that the data may not be normally distributed. The NO2 graph shows a similar trend for the fracturing phase. We did not observe significant differences for different development phases for ozone, which is not surprising as it is a secondary pollutant and it can be related to the winter season of the data collection period (Edwards et al., 2014). There was a dramatic increase in ethane concentration for the flowback phase. This 25th percentile was 24 ppb, while this concentration ranged between 0 and 11 ppb for other phases. The 75th percentile was 89 ppb, which is a significantly higher value compared to other phases. We observed a similar trend for methane concentration. The 25th percentile (2.5 ppm) and the 75th percentile (4.3 ppm) were significantly higher than other phases. Differences for development phases for CO2 were not statistically significantly different. CO2 has many emissions sources and variable background concentrations, so distinguishing emissions from the well pad activities is difficult.

The average concentrations of methane and ethane for the entire monitoring campaign are shown in Fig. 4a. The highest ethane concentrations occurred during the flowback stage (565.7 ppb). A mean that is significantly higher than the median comes from a distribution that is skewed due to peak events (meanethane=11.4 ppb, medianethane=8.5 ppb). Figure 4b shows the time series of ethane-to-methane ratios throughout the operational phases. The lowest average ratio occurred at horizontal drilling with 2.5, while the highest ratio occurred at the flowback phase with an average ratio of 17.4. The average ethane / methane ratios for fracturing, drillout, and production phases are 3.4, 3.2, and 5.1, respectively.

The hourly concentration graphs of NOx, O3, CH4, and CO2 were used to analyze the factor solutions (Fig. S8). Hecobian et al. (2019) investigated the emissions during different well pad development phases to analyze emission rates in the Denver-Julesburg and Piceance basins in Colorado, US. They observed that emission rates of benzene and most VOCs were highest during flowback for both basins; on the other hand, they had much lower emission rates from the production phase, which can be related to the differences in duration of each phase (days to weeks). Light alkanes and benzene concentrations were higher during hydraulic fracturing. It is difficult to directly compare the results of the two studies, because the proposed study is based on continuous data during each phase while Hecobian et al. (2019) collected 374 measurements from five drilling, eight fracking, nine flowback, one liquid load out, and 11 production sites to analyze emission rate.

Figure S4 shows the dominant wind directions on overall concentrations, as well as giving information on the different concentration levels. Pollution roses show which wind directions contribute most to overall mean concentrations. For all air quality species, southwestern winds control the overall mean concentrations at the well pad. To explore the relationship between methane and ethane, we conditioned ethane by methane. Figure S5 indicates that higher ethane concentrations are associated with the SW and higher methane concentrations. The results also show that lower ethane and methane concentrations contributed from the east; the highest methane concentrations were obscured by a relatively high ethane background. The highest contribution to the factors was provided by the SW data (Fig. S6).

The three-factor model was chosen for the PMF analysis based on the interpretation of the factor profiles, Qrobust / Qexpected ratios (Table S3), factor contributions, error estimation results (Table S4, Fig. S9), and hourly peak concentrations of pollutants. The three-factor solution was resolved to the following factors: natural gas for the natural-gas-related emissions sources, regional transport/photochemistry for the atmospheric regional molecular transport and oxidized background air, and engine emissions for emissions from vehicles, drill rigs, generators, and pumps used at the site (Fig. 5). The summary of PMF models with various peak values for well development activities are shown in Table S4. The DISP, BS, and BS–DISP results for two-, three-, and four-factor PMF solutions are summarized in Table S2. For the three-factor analysis, the DISP results indicate that there are no swaps and the PMF solution is stable, which means there are no exchange factor identities and it is a well-defined solution for the case. According to BS results, there is a small uncertainty; this could be an impact of high variability in concentration. BS–DISP captures both random errors and rotational ambiguity; these results also indicate that the solution is reliable because there are no swaps between factors for the PMF model. Error estimation summary plots (Fig. S9) show a range of concentration by species in each factor: base value, BS 5th, BS median, BS 95th, BS–DISP 5th, BS–DISP average, BS–DISP 95th, DISP min, DISP average, and DISP max.

The natural gas factor was named as such due to its composition of species that are present in natural gas: 89 % CO2, 1 % methane, 3 % ethane, 1.5 % propane, 0.5 % isobutane, 1 % n-butane, 0.1 % pentane, and 0.2 % isopentane (Fig. S10). Ethane is a particularly good marker for natural gas emissions sources because its atmospheric sources are almost exclusively from natural gas extraction, production, processing, and use (Liao et al., 2017). A total of 92 % of ethane mass is explained by the natural gas factor (Fig. 6). The highest contribution for this factor occurred during the flowback phase.

The regional transport/photochemistry factor was characterized by high contributions from ozone (12 %), methane (1 %), and CO2 (86 %) (Fig. S10). A total of 99 % of the ozone mass was explained by this factor (Fig. 6). Ozone is a product of photochemistry and not directly emitted by any of the sources on the well pad. Although CH4 and CO2 would be emitted by well pad sources, they are also present in background ambient air and could be transported to the monitoring location from other sources in the region. Contributions of this factor were relatively steady for all phases of operation during the entire monitoring campaign.

The engine emissions factor was composed of 39 % NOx, 33 % NO, and 11 % NO2 as well as 0.02 % toluene and 0.04 % benzene (Fig. S10). The portions of the mass of these species explained by this factor are 74 %, 87 %, 60 %, 20 %, and 54 %, respectively (Fig. 6). Toluene is released mainly from motor vehicle emissions and chemical spills (Gierczak et al., 2017). Another important emission source is oil and gas extraction (EPA, 1993). Contribution of this factor was significantly highest during hydraulic fracturing, when there were emissions from many diesel engines operating continuously on the well pad. Contribution during flowback was also elevated. Several peaks of contribution were observed during production, which could be due to maintenance vehicles and other short-lived vehicle-based activities on the well pad.

The main limitation of the study is having an uneven number of data points for each operational phase. This limitation affects the analyses; however, we do not have control of the durations of the operational phases. PMF models have several limitations. First, they need large datasets. In this study, the number of data varies based on the duration of the activity (Fig. S2). Therefore, the contribution to the factors is not the same for each phase. This is the main reason behind the uncertainty of defined factors. Second, the accuracy and precision of measured species limit the analysis. The determination of the number and character of factors is based on an expert's interpretation. Comprehensive information is needed on source profiles to verify the defined source profiles. Finally, the pre-set parameters play an important role in the model results. For future work, integrating more data from different fields can decrease the inherent uncertainty.